lateral-thinking: a skill pack that teaches your agent to abandon bad ideas

The demo is the pitch

The README leads with a side-by-side that makes the argument better than any feature list. Same agent, same prompt about users who create one project and never return. The vanilla answer is three sensible features: an onboarding checklist, a weekly email, a personalized dashboard. The answer after reading the random-stimulus skill produces different product directions: a lighthouse-inspired beacon that signals only when a user's own project hits a date, risk or decision, and a sourdough-starter-inspired pulse that turns the project into something alive between visits, with a Ferris-wheel idea explicitly abandoned for being a prettier dashboard.

The README names the meta-pattern the techniques exist to surface: retention is not a reminder problem, it is a new-signal problem, and three features that all ask the user to remember the product are one idea in three costumes. For anyone who has watched an agent brainstorm, the failure being addressed is recognizable: fluent lists where every item shares a skeleton.

Installing the toolkit

One command covers every agent that reads skills:

npx skills add danium/lateral-thinking

Claude Code users get the plugin route:

/plugin marketplace add danium/lateral-thinking
/plugin install lateral-thinking@lateral-thinking

Codex users ask the skill-installer to pull from the GitHub URL, and the manual path is cloning and copying the skills into your agent's directory. One install note carries real information: the toolkit installs as a bundle because the lateral router needs its sibling techniques. That dependency is the design in miniature; the router is a diagnostician, and without the eight specialists to route to, its diagnosis would be advice with nowhere to go.

A router that diagnoses the stuck

The routing table maps symptoms to techniques, and the symptoms are written from inside real meetings. Ideas all feel the same routes to random-stimulus. A constraint feels unbreakable routes to provocation. Requirements assume things nobody questioned routes to inversion. We might be solving the wrong problem routes to concept-fan; a working but derivative solution to analogy; one idea needing variations to scamper; a decision being made too fast, or everyone agreeing too easily, to six-hats; and everything feeling timid to worst-idea.

An applied skill named post rounds out the set, taking a topic or someone's post and routing it through the toolkit into drafts you can send. The README also manages expectations about invocation: auto-invocation is best-effort and agent-dependent, and the reliable path is naming the skill or asking your agent to use the router.

Failure as a feature

The section that lifts this above a prompt pack is the anatomy of a session, where the scamper technique is shown refusing to pad its own output mid-run. Three of its seven operations come up empty, stated plainly with reasons: the command already has its substitution, the install generates no by-product worth repurposing, and the one genuine reversal was already counted as another variant, so it was disqualified as a duplicate.

The README draws the right conclusion from that transcript. An agent that produces seven strong ideas from seven prompts is pattern-matching the shape of a good answer, and you cannot tell which ideas it actually believes. Every skill here carries an abandonment rule, a stated condition for giving up, and an instruction to show you when it fires. The visible failures are what make the surviving ideas credible, which is an honesty mechanic more ideation tools could steal.

Provenance, from de Bono to the d.school

The techniques are sourced, and the sources are the canon. Random stimulus, provocation and the concept fan come from Edward de Bono's lateral thinking books across 1970 to 1992. Six Thinking Hats is de Bono 1985, with a note that it is a registered trademark of the de Bono Group and this is an independent educational implementation. SCAMPER is credited to Bob Eberle building on Alex Osborn's checklist, forced analogy to Synectics, assumption inversion to classical dialectic with the Jacobi via Munger phrasing, and worst possible idea to the reverse-brainstorming standard in the d.school and IDEO toolkit.

The provenance matters for two reasons. It means each skill encodes a method with decades of field testing rather than a model's improvisation, and the contributing rules require the same for additions: new techniques need provenance and honesty mechanics, and worked examples must come from real sessions.

Limits, honestly placed

The constraints are stated in the README rather than discovered by users. Auto-invocation is unreliable, so plan on explicit requests. The toolkit shapes ideation, not implementation; nothing here writes the feature after the direction is chosen. The techniques constrain how an agent generates and the honesty mechanics constrain what it may claim, but the README is careful to say this is not magic, structured divergence with quality rules attached.

Project facts are modest: MIT licensed, no releases or tags, and the last push on 2026-09-13, so it is a young, fast-moving artifact. The value proposition is also the maintenance burden: abandonment rules and worked examples from real sessions are expensive to write and cheap to fake, which makes the contributing requirements the load-bearing wall of the whole repository's credibility.

Against prompt packs and brainstorm frameworks

The alternatives are easy to enumerate. A well-written brainstorm prompt gets you the vanilla column of the demo: sensible, same-shaped output. Books on de Bono's methods and design-school toolkits contain the same techniques with more depth, but they live outside the agent, and the whole premise of this pack is that divergence methods only help if they run where the ideas are generated. Generic ideation skills exist in the agent ecosystem; the differentiators here are the router, the abandonment rules, and the sourced provenance.

The honest boundary is that no technique pack guarantees a good direction; it guarantees that the agent's process is visible, falsifiable and varied. For a product person who regularly asks an agent the stuck-question, that visibility converts brainstorming from a slot pull into a reviewable method, and eight named techniques with failure states is a reasonable unit of purchase.

Editorial conclusion

lateral-thinking fits product people and engineers who brainstorm with coding agents, have noticed the same-shaped output problem, and will actually invoke the router rather than hoping for auto-invocation. Skip it if you want a creativity oracle rather than a method, or if your workflow never asks agents for directions in the first place. Verify the premise in one session: install it, rerun a real stuck-question from your product through random-stimulus, and compare the abandoned ideas against the delivered ones to see whether the visibility is worth the ceremony.

Frequently asked questions

Official sources

1. danium/lateral-thinking on GitHub
2. Issues
3. License: MIT
4. README